159 results about "Gradient estimation" patented technology

Systems and techniques are described herein for training a machine learning (ML) model. For instance, a process can include obtaining, from a teacher ML model, a first prediction based on an input. The process can further include obtaining, from a student ML model, a second prediction based on the input. The process can include determining a loss based on a difference between the second prediction from the first prediction. For instance, the loss can include a variance reduced total variation distance (TVD) loss based on an unbiased gradient estimate of the loss. The process can further include backpropagating the loss through the student ML model to train the student ML model.

The invention belongs to the technical field of image processing, and particularly relates to a road segmentation and gradient estimation method based on multi-sensor fusion. Comprising the following steps: S1, registering a camera and a laser radar; s2, generating a road segmentation result based on output data of the camera and the laser radar after registration; and S3, obtaining a three-dimensional point cloud image of the road surface based on the road segmentation result, and realizing slope estimation of the road surface based on the three-dimensional point cloud image. According to the method, the error rate of road surface segmentation and the error of slope estimation are effectively reduced through complementation of multiple sensors, and stable output is still kept under complex backgrounds such as vibration, rain and fog and weak light.

The invention discloses an automatic driving test scene set optimization method, apparatus and device, and a storage medium. The method comprises the steps of generating a simulation scene file in an OpenSCENARIO format through preprocessing data; analyzing the simulation scene file through a teacher model, outputting a risk description text, receiving a scene feature vector and the risk description text through a student model, and outputting a failure probability; determining a comprehensive value index according to the failure probability, dynamically updating a scene library of automatic driving test scenes, collecting failure data in an AUT test, performing incremental fine tuning on the student model, and obtaining an optimized target test scene set; the problem of gradient estimation variance explosion caused by sparseness disasters can be effectively solved, the stability of model training is improved, the recognition capability of a rare long-tail failure scene is enhanced, the accuracy of failure probability prediction is improved, the judgment accuracy of the model to a boundary scene is improved, and the accuracy and comprehensiveness of scene value quantification are improved.

The invention relates to the technical field of road repair state monitoring, in particular to a road repair state monitoring method and system based on edge calculation. The method comprises the following steps: acquiring planning information of a repaired road, deploying 3D laser scanning vibration meters at a plurality of repairing points, collecting pavement vibration signals in different time periods, and performing disordered distribution time domain feature analysis to obtain a vibration signal disordered distribution time domain graph; then, structure weakening behavior simulation is carried out based on the graph, periodic structure weakening data is quantified, and a pavement water damage cumulative gradient is obtained through water damage cumulative gradient estimation; and finally, in combination with the periodic structure weakening data and the water damage cumulative gradient, carrying out service life evaluation on the restoration state, and sending an evaluation result to the terminal. According to the invention, the road repair state monitoring technology is optimized, so that the road repair state monitoring technology is more perfect.

The invention discloses a fine tuning system of a large-scale pre-training model in a federated learning environment and application thereof. The system comprises a local disturbance gradient estimation module, a differential privacy protection module and a global model aggregation and update module. The local disturbance gradient estimation module is used for calculating a global model loss value by combining forward propagation with a zero-order optimization method so as to estimate a gradient and realize all-parameter fine tuning; the differential privacy protection module performs differential privacy protection processing on the estimated disturbance gradient to prevent gradient information from leaking user sensitive data; and the global model aggregation and update module reconstructs a disturbance vector and completes global model update based on a random seed and a scalar gradient uploaded by a client. Compared with the prior art, on the premise of not depending on back propagation, all-parameter fine tuning of a large-scale pre-training model is achieved, data privacy is guaranteed, meanwhile, calculation and memory expenses are remarkably reduced, and the method is suitable for a resource-limited distributed calculation environment.

The invention discloses an automatic thinking chain prompt generation method based on black box optimization and vulnerability quantification. The method comprises the following steps: acquiring an unlabeled training data set; screening a high-difficulty problem subset through a difficulty judgment module, and generating an annotation data set; generating a plurality of reasoning chains for each question in the annotation data set by using a large language model, and reserving the reasoning chains consistent with correct answers to form an example library; a variance reduction strategy gradient estimator is adopted to optimize an inference chain selection strategy, and a common style prompt is generated; and performing style diversification processing on the common style prompt through a novel format prompt module to generate an anti-vulnerability automatic thinking chain prompt.

The invention discloses a parameter determination method and device of an analog circuit, a medium and a product, and relates to the technical field of integrated circuit design. Local performance feedback and global topology perception of performance index parameters are used as input parameters, and pertinence and generalization ability of parameter adjustment are remarkably improved. Through a strategy function, a parameter candidate set of a target analog circuit is obtained, through multi-group sampling processing, variance of strategy gradient estimation is reduced, and training stability and convergence speed are improved. Depending on a value network is not needed, and training resource overhead is reduced. And applying the parameter candidate set output by the strategy function to the simulation processing scene to realize the accuracy of simulation processing evaluation. The target performance indexes are extracted from the simulation result, optimization of the performance indexes is achieved, the multiple target performance indexes are processed according to the multi-target reward function, balance selection and rejection are achieved in the multiple performance indexes, the parameter combination corresponding to the highest advantage value is obtained, and the reliability of the analog circuit parameter determination process is improved.

The invention discloses a multi-robot fixed time cooperative hunting method and system based on a distributed time-varying optimization algorithm. The method comprises the following steps: establishing a dynamic model and a communication topological graph of a multi-robot system; the method comprises the following steps: converting a fixed time distributed hunting problem of a multi-robot system into a fixed time distributed time-varying optimization problem, and constructing a global objective function; designing a fixed time distributed gradient estimator, and enabling each robot to estimate gradient information of a system global objective function within fixed time in a distributed mode; designing a self-adaptive zero-order neural network which is used for approximating an inverse matrix of a Hessian matrix; and designing a fixed-time distributed time-varying optimization algorithm for the multi-robot system by adopting the gradient information of the global objective function estimated by the distributed gradient estimator in the step 3 and the inverse matrix of the Hessian matrix approximated by the adaptive zero-order neural network in the step 4, so that each robot encircles the dynamic target in the fixed time in a distributed mode. According to the invention, the multi-robot system is ensured to surround the moving target in a formation form within a fixed time.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a neural network by estimating the objective function curvature based on current and previous gradients. In one aspect, a method comprises: sampling a batch of training data; and for each neural network parameter: determining, based on the current batch of training data, a respective current gradient of the objective function at the current iteration with respect to the current neural network parameter; estimating an objective function curvature with respect to the current neural network parameter based on (i) the current gradient of the objective function at the current iteration, and (ii) a respective previous gradient of the objective function at each of a plurality of previous iterations; and updating a current value of the neural network parameter based on the estimate of the curvature of the objective function.

The invention discloses a control method of a flat wire motor. The method comprises the following steps: constructing an optimal current instruction table taking a rotating speed and a torque as indexes; if the motor is in the dynamic working condition, directly looking up a table to obtain a current optimal current instruction; if in a steady-state working condition, performing online search by taking a table look-up result as an initial point: generating two-dimensional Bernoulli random disturbance, and sequentially applying positive and negative disturbances to the initial point; respectively measuring the input power under the two disturbances; according to the ratio of the positive and negative input power difference to the corresponding disturbance component, synchronously calculating gradient estimation vectors of the system input power to d-axis and q-axis current; according to a steepest descent method, the current instruction is updated in the reverse direction of the gradient, an optimal instruction enabling the input power to be minimized is obtained, and an inverter is driven to control a motor to operate. According to the method, through a hybrid control strategy combining dynamic table look-up and steady-state online search, the total loss minimization of the flat wire motor under all working conditions in consideration of alternating current copper loss is realized, and the rapidity of dynamic response is ensured.

The invention discloses a slope estimation method and device, equipment and a storage medium. The method comprises the steps that the current corresponding driving working condition of a vehicle is obtained; target limiting parameters matched with the driving working conditions are determined; performing limiting processing on the original acceleration difference value by using the target limiting parameter to obtain a target acceleration difference value; the original acceleration difference value is the difference value between the detection acceleration and the motion acceleration of the vehicle, the detection acceleration is obtained through detection of an acceleration sensor arranged on the vehicle, and the motion acceleration is determined based on the vehicle speed; and estimating the gradient of the road where the vehicle is currently located by using the target acceleration difference value. In this way, the accuracy of road slope estimation can be improved.

The embodiment of the invention provides a large language model optimization method based on variance reduction and momentum acceleration. The method comprises the following steps: a gradient estimation stage of a large language model: initializing a seed list and a projection list; executing multiple independent query iterations, calling a disturbance subprogram, generating a random seed for the large language model, and storing a table; in the disturbance subprogram, the determined gradient projection value is stored in a projection list, and iteration is carried out for next query; after executing multiple independent query iterations, storing a plurality of random seeds and a plurality of gradient projection values corresponding to the random seeds; in the weight updating stage of the large language model, a gradient norm subprogram is called for each layer of the large language model, and a random seed is obtained to reset a random number generator; and determining variance-reduced gradient estimation according to the gradient projection value extracted from the projection list and the reproduced disturbance vector. According to the embodiment of the invention, gradient information queried for multiple times is aggregated to generate low-noise gradient estimation, and fine tuning of a large language model is completed.

The invention relates to the field of generative artificial intelligence image processing, and provides an initial noise distribution optimization method, device, equipment, medium and program, and the method comprises the steps: taking a denoising process as a fixed mapping relation, creating a trainable distribution parameter, and constructing a distribution parameter updating formula based on the fixed mapping relation; constructing a dynamic reward calibration module, calculating a difference value between a reward value of a current initial noise distribution generated image and a reward value of an original standard normal distribution generated image after the diffusion model outputs the generated image every time, and taking the difference value as a relative reward value; constraining the updating process of the distribution parameters by adopting a proportional clipping algorithm; and calculating a parameter updating step length based on a gradient estimation result, and carrying out proportional clipping on the updating step length. The method is used for improving the consistency of content and prompt semantics in a text-to-image generation task by optimizing the initial distribution parameters of the diffusion model, and meanwhile, the generation quality and the calculation efficiency are kept.

The invention discloses a robot group detection scheduling method and system, and relates to disaster monitoring: scene reconstruction is carried out according to collected real-time sensing information, a three-dimensional digital model of a disaster scene is generated, and a discrete monitoring grid based on sampling points is established; constructing a disaster risk assessment model on the discrete monitoring points, and for gas and fire, calculating the risk value of each monitoring point by using actual measurement data to form a discrete risk distribution map; calculating the space change trend of the risk degree by adopting an adjacent point difference method, and calculating a directed gradient through the risk degree difference value and the distance of adjacent monitoring points; when the risk degrees of any two disasters are increased at the same time, marking the risk points as potential coupling risk points; identifying a propagation path of the risk degree based on connectivity analysis of the monitoring network; and scheduling the robot group by adopting a risk avoiding path planning algorithm. According to the method, the spatial distribution and evolution trend of the disaster field are accurately reconstructed through local gradient estimation under the sparse sampling condition.

The invention belongs to the technical field of neural network model training optimization, and discloses a low-rank matrix gradient estimation method and system for large-scale neural network training, and the method comprises the steps: sampling a low-rank random subspace meeting the equidistant or isotropic constraint based on a preset sampling rule; a matrix gradient estimation algorithm is embedded into the low-rank random subspace to be executed, and accumulation and updating are carried out on low-dimensional auxiliary variables; when the low-dimensional auxiliary variable is accumulated to a preset step number, performing inertia updating on parameters of a matrix gradient estimation algorithm; and calculating a weighting matrix based on all parameters of the matrix gradient estimation algorithm after inert updating, and optimizing a preset sampling rule of a next round of low-rank random subspace according to spectrum information of the weighting matrix so as to realize gradient estimation of the low-rank matrix. According to the method, a scheme including estimation stage low-rank, projection distribution optimization and inertia updating is provided, the video memory and step time threshold is remarkably reduced in engineering, and considerable cost performance is embodied in large model fine tuning.

Systems and techniques are described herein for training a machine learning (ML) model. For instance, a process can include obtaining, from a teacher ML model, a first prediction based on an input. The process can further include obtaining, from a student ML model, a second prediction based on the input. The process can include determining a loss based on a difference between the second prediction from the first prediction. For instance, the loss can include a variance reduced total variation distance (TVD) loss based on an unbiased gradient estimate of the loss. The process can further include backpropagating the loss through the student ML model to train the student ML model.

The invention discloses a fine tuning system of a large-scale pre-training model in a federated learning environment and application thereof. The system comprises a local disturbance gradient estimation module, a differential privacy protection module and a global model aggregation and update module. The local disturbance gradient estimation module is used for calculating a global model loss value by combining forward propagation with a zero-order optimization method so as to estimate a gradient and realize all-parameter fine tuning; the differential privacy protection module performs differential privacy protection processing on the estimated disturbance gradient to prevent gradient information from leaking user sensitive data; and the global model aggregation and update module reconstructs a disturbance vector and completes global model update based on a random seed and a scalar gradient uploaded by a client. Compared with the prior art, on the premise of not depending on back propagation, all-parameter fine tuning of a large-scale pre-training model is achieved, data privacy is guaranteed, meanwhile, calculation and memory expenses are remarkably reduced, and the method is suitable for a resource-limited distributed calculation environment.

The invention discloses a data-free federated distillation method and system based on zero-order gradient estimation, and the method comprises the steps: initializing a global model and a synthetic image generator through a central server, generating a synthetic image sample in each communication round, transmitting the synthetic image sample to a client, and enabling the client to update a local model through local privacy data, and uploading a prediction result. And the central server calculates the gradient of the generator by using a zero-order gradient estimation technology, and updates generator parameters and global model parameters. And by introducing fidelity loss, adversarial loss, diversity loss and negative information entropy loss, the performance of the generator is optimized. And the gradient of the generator is calculated through zero-order gradient estimation, so that the requirement of accessing a local model of a client is avoided, and privacy is effectively protected. The global model is updated by minimizing knowledge distillation loss, and at the same time, the client further optimizes the local model by receiving an integrated prediction result. The method has the advantage that the communication bandwidth requirement is reduced by reducing the access to the private data of the client.

A control device 50 as a road gradient estimation device estimates the road gradient of a road on which a vehicle 10 is traveling on the basis of an inertial acceleration derived on the basis of an inertial force in a traveling direction of the vehicle 10 and an actual acceleration derived on the basis of the amount of change in the vehicle speed of the vehicle 10. The control device 50 includes a forward / backward movement determination unit M11 for modifying a forward / backward movement determination indicating whether the vehicle 10 is moving backward. The control device 50 includes an acceleration correction unit M12 that acquires the inertial acceleration as a first acceleration and the actual acceleration as a second acceleration, and, when the forward / backward movement determination is modified, performs correction to invert the sign of the first acceleration or the second acceleration. The control device 50 includes a gradient estimation unit for estimating the road gradient on the basis of the difference between the first acceleration and the second acceleration.

The invention relates to the field of artificial intelligence security, and provides an adversarial sample optimization method and device, equipment and a program product. The method comprises the following steps: acquiring an audio sample; according to the audio sample, obtaining a candidate sample by using a preset genetic algorithm; and according to the candidate sample, performing sparse gradient estimation in combination with the target model, and optimizing the candidate sample in combination with a sparse gradient estimation result. According to the confrontation sample optimization method provided by the invention, the acquired audio sample is globally searched by using the preset genetic algorithm to find the closest candidate sample and avoid falling into a locally optimal solution, and sparse gradient estimation is combined to more accurately capture the key decision boundary of the target model, so that a more effective confrontation sample can be generated, and the accuracy of the confrontation sample optimization is improved. Meanwhile, unnecessary disturbance is avoided, the optimization pertinence is improved, the real-time requirement is met, the calculation overhead is reduced, and the optimization efficiency is improved.

A road surface gradient estimation device includes an acquisition unit that acquires each of detection results of an acceleration detection unit that detects an acceleration in a front-rear direction of a vehicle, and a wheel speed detection unit, and acquires information regarding power for driving the vehicle, a derivation unit that derives a gradient of a road surface on which the vehicle is traveling as a first road surface gradient based on the acquired acceleration and wheel speed, and a correction unit that corrects the derived first road surface gradient based on the information regarding the power for driving the vehicle to derive a second road surface gradient.

An image recognition method of a pulse neural network based on Fourier and cross-time constraints is characterized by comprising the following steps: constructing an image recognition system of the pulse neural network based on Fourier and cross-time constraints; the image acquisition module acquires an image sample set, and pre-processes the image sample set to obtain a training set; training the network by using a training set, enabling pulse neurons to emit pulses by using a step function in forward propagation, and adopting a derivative of finite term Fourier series as gradient estimation of the pulses on membrane potential in reverse propagation; measuring the similarity between the time steps by using cosine similarity, calculating the loss of the target function of the network after the rth training, and adjusting network weight parameters according to the loss of the target function; and carrying out image classification identification operation on the preprocessed to-be-identified image data by using the trained network, and outputting an image classification result. The method has the effect of improving model performance.

The present invention proposes an adversarial sample generation method and system based on dynamic advanced iteration, belonging to the field of artificial intelligence security technology. The method includes: preprocessing the original image to generate an initial adversarial sample; initializing the running parameters, inputting the running parameters and the initial adversarial sample into a surrogate model for iterative calculation, calculating an advanced factor before each iteration, judging the stage of the iterative process according to the current iteration number, and dynamically adjusting the advanced factor; calculating the advanced position and the gradient of the loss function based on the cumulative momentum and the dynamically adjusted advanced factor, updating the cumulative momentum based on the gradient of the loss function, and generating an adversarial sample, and obtaining the finally generated adversarial perturbation by using a clipping function; when the maximum iteration number is reached, adding the finally generated adversarial perturbation to the original image to generate a final adversarial sample; if not, repeating the above steps. It effectively solves the problems of the limitation of the fixed offset, the momentum stability, and the accuracy of gradient estimation.

The invention provides a slope estimation method and system for all working conditions of an automobile, and relates to the technical field of automobile control. The slope estimation method for the full working condition of the automobile comprises the steps of obtaining automobile state information; obtaining dynamic ramp estimation data based on a preset dynamic ramp estimation algorithm and the automobile state information; obtaining kinematic ramp estimation data based on a preset kinematic ramp estimation algorithm and the automobile state information; dividing automobile working conditions based on the automobile state information to obtain automobile working condition data; and performing data fusion according to the dynamic ramp estimation data, the kinematic ramp estimation data and the automobile working condition data to obtain a gradient estimation result. According to the slope estimation method for the full working condition of the automobile, the technical effect of improving the accuracy and economical efficiency of slope estimation under the full working condition can be achieved.

A system, method, and computer program product for training a generative language model (GLM) is provided. A plurality of sampled rationales for various question-answer pairs are generated using the GLM. A gradient estimate of parameters of neurons in the GLM is determined based on these sampled rationales to maximize the learning objective of the GLM. The parameters of the GLM are modified using the gradient estimate over multiple iterations, ultimately providing a trained GLM.

To reduce the processing load required to recognize the external environment surrounding the vehicle. [Solution] The external environment recognition device 50 includes an on-board detector 5, a recognition unit 111 that recognizes the road surface and three-dimensional objects on the road on which the vehicle is traveling as road surface information based on point cloud data for each frame acquired by the on-board detector 5, a determination unit 113 that determines the interval of detection points necessary for the point cloud data of the next frame based on the size of a predetermined three-dimensional object set in advance as a recognition target and the measured distance from the vehicle to the three-dimensional object based on the point cloud data, and a gradient prediction unit 114 that predicts the gradient of the road surface not recognized by the recognition unit 111 based on gradient information associated with map information on which the road is recorded. The determination unit 113 further determines the interval of detection points necessary for the point cloud data of the next frame based on the size of a predetermined three-dimensional object, the map information and the estimated distance from the vehicle to the three-dimensional object estimated from the gradient, for a range from the furthest distance to the required distance.

The application discloses a black box optimization gradient estimation method and device based on Rung-Kutta extrapolation and a medium, the method of which comprises the following steps: obtaining a target function to be optimized and a gradient solving reference point corresponding to the target function; configuring gradient calculation related parameters, generating a multi-scale step set and a unit orthogonal perturbation vector set based on the gradient calculation related parameters; determining perturbation positions of the target function under different steps based on the gradient solving reference point, the multi-scale step set and the unit orthogonal perturbation vector set, and converting response data corresponding to each perturbation position into numerical differential data; weighting and fusing the numerical differential data corresponding to different scale steps based on a preset fusion weight; then performing correlation processing on each unit orthogonal perturbation vector to obtain multiple gradient components, and obtaining a final gradient estimation result after aggregation processing of each gradient component. According to the application, the Rung-Kutta extrapolation technology is used to systematically offset low-order truncation errors, and the gradient estimation precision is significantly improved.

The present invention discloses a cross-device federated learning method for minimum-maximum problems, and belongs to the technical field of federated learning. The method includes: a central server initializes the principal variables and dual variables in the model parameters, as well as the number of iterations; the central server selects a subset of clients and sends the model parameters to each client; each client in the client subset calculates a local gradient estimator; the central server receives the local gradient estimator returned by the client and calculates a global gradient estimator; the central server selects another subset of clients, sends the model parameters and the global gradient estimator to each client, and the client performs a K-step local update on the model parameters and sends the final local model parameters to the central server; after the central server receives the local model parameters returned by the client, it calculates new global model parameters, iterates the calculation, and outputs the final parameters.

The application relates to a fast wideband signal detection method, device and equipment based on a time-frequency graph and a medium. The method determines the starting and ending frequencies of a wideband signal by performing Gaussian blurring, time domain averaging, gradient estimation and gradient matching on a time-frequency graph of an electromagnetic signal; verifies and scores the starting and ending frequency positions of each row on the time-frequency graph, and if the score passes, the row contains a wideband signal, otherwise, the row does not contain a wideband signal; determines whether the wideband signal is a continuous signal, and if it is not, calculates the starting and ending times of a wideband signal pulse in the time-frequency graph according to the number of rows containing the wideband signal; and obtains the characteristics of the wideband signal in the time-frequency graph according to the starting and ending frequencies of the wideband signal and the starting and ending times of the wideband signal pulse. The method proposes different feature extraction processes and detection decision mechanisms, guarantees the timeliness of detection, and provides theoretical support for fast wideband signal detection in complex environments.

The invention relates to an automatic focus following method for a microscope, relates to the technical field of microscope imaging, and aims to construct a defocus distance prediction model by taking focusing as a regression problem so as to directly predict a defocus distance based on a picture. In the initial stage of focusing search, a gradient estimation algorithm without global scanning is provided based on a defocus distance prediction model, and the optimal focal plane is quickly converged through the adjustment steps of detecting a current area, predicting a focusing position, jumping to the focusing position and using the least imaging times and the least camera height, so that the focusing time is remarkably shortened, and the focusing efficiency is improved. The technical problems that an existing focusing algorithm is large in operand, and focusing to a wrong position is prone to occurring are solved.